Deep Compressed Sensing Yan Wu, Mihaela Rosca, Tim Lillicrap - PowerPoint PPT Presentation

Deep Compressed Sensing Yan Wu, Mihaela Rosca, Tim Lillicrap

Compressed Sensing A Brief Review An underdetermined problem: random projection signal, e.g, vectorised image measurements M x y Reconstruction of x is possible when the signal is sparse (Candes, Donoho, Romberg, Tao, 2006~) : Deep Compressed Sensing — Yan Wu

Why CS works Restricted Isometry Property (RIP/REC) The projection M preserves the Euclidean distance between k-sparse signals ● ● Many random matrices have the RIP with high probability Examples: MRI reconstruction, the single pixel camera Deep Compressed Sensing — Yan Wu

RIP can be a trained property MNIST Baseline: Compressed Sensing using Generative Models (Bora et al. 2017, almost the same as our model except using CelebA separately trained generators) Deep Compressed Sensing — Yan Wu

Improve GANs by online optimisation CIFAR, Deep Convolutional GAN DCGANs, sweeps over 144 hyper-parameters: Spectral-normalised GANs Deep Compressed Sensing — Yan Wu

Summary ● A framework based on minimising measurement errors ● Bring RIP to neural networks via training ● Improve GANs: novel use of the discriminator as a measurement function ● A new semi-supervised model Model Metric Property Compressed Sensing RIP from random projection Deep Compressed Sensing Trained RIP Semi-supervised GANs Multi-Class Classifier CS-GANs Binary Classifier … … Poster #24, Pacific Ballroom Deep Compressed Sensing — Yan Wu